Transcript Slide 1
Natural Language Processing for Information Access Horacio Saggion Department of Computer Science University of Sheffield England, United Kingdom [email protected] Overview of the course Introduction and Motivation (Day 1) NLP technology and tools (Day 1 & 2) Question Answering (Day 3 & 4) Text Summarization (Day 4 & 5) Outline NLP for Information Access Components tokenisation Information Retrieval sentence splitting Information Extraction part of speech tagging Text Summarization named entity recognition Question Answering morphological analysis Cubreporter: a case study parsing Other applications Demonstrations GATE tools for NLP Information Retrieval (Salton’88) Given a document collection and a user information needs Produces lists of documents matching the information need Information needs can be expressed as sets of keywords Documents are pre-processed in order to produce term indexes which contain information about where each term occurs in the collection - decisions have to be taken with regards to the definition of term Information Retrieval Needs a method to measure the similarity between documents and queries The user has to read the documents in order to find the desired information IR can be applied to any domain Text Retrieval Conferences (since 1992) contributed to system development and evaluation (http://trec.nist.gov) Information Extraction (Grishman’97) Pulls facts from the document collection Based on the idea of scenario template some domains can be represented in the form of one or more templates templates contain slots representing semantic information IE instantiates the slots with values IE is domain dependent – a template has to be defined Message Understanding Conferences 1987-1997 fuelled the IE field and made possible advances in techniques such as Named Entity Recognition From 2000 the Automatic Content Extraction (ACE) Programme Information Extraction ALGIERS, May 22 (AFP) - At least 538 people were killed and 4,638 injured when a powerful earthquake struck northern Algeria late Wednesday, according to the latest official toll, with the number of casualties set to rise further ... The epicentre of the quake, which measured 5.2 on the Richter scale, was located at Thenia, about 60 kilometres (40 miles) east of Algiers, ... DATE DEATH INJURED 21/05/2003 538 4,638 EPICENTER Thenia, Algeria INTENSITY 5.2, Ritcher Information Extraction Template can be used to populate a data base Template can be used to generate a short summary of the input text A 5.3 intensity earthquake in Algeria killed more than 500 people. Data base can be used to perform reasoning What Algerian earthquake killed more people? Information Extraction Tasks Named Entity recognition (NE) Coreference Resolution (CO) Identifies identity relations between entities in texts Template Element construction (TE) Finds and classifies names in text Adds descriptive information to NE results Scenario Template production (ST) Instantiate scenarios using TEs Examples NE: CO: a powerful earthquake and the quake ALGIERS and Algiers TE: Thenia (Location), Algiers (Location), May 22 (Date), Wednesday (Date), etc. entity descriptions: Thenia is in Algeria, Algiers is capital of Algeria ST combine entities in one scenario (as shown in the example) Question Answering (Hirschman&Gaizauskas’01) Given a document collection (can be the Web) and a natural language question Extract the answer from the document collection answer can in principle be any expression but in many cases questions ask for specific types of information such as person names, location, dates, etc. Open domain in general Text Retrieval Conferences Question Answering Track responsible for advances in the field of system development and evaluation (since 1999) From 2008 the Text Analysis Conference QA Task (Voorhees’99) In the Text Retrieval Conferences (TREC) Question Answering evaluation, 3 types of questions are identified Factoid questions such as: List questions such as: “Who is Tom Cruise married to?” “What countries have atomic bombs?” Definition questions such as: “Who is Aaron Copland?” or “What is aspirin?” (Changed name to “other” question type) Text Summarization (Mani’01) Given a document or set of documents Extract the most important content from it and present a condensed version of it extracts vs abstracts Useful for decision making: read or not read; saves time; can be used to create surrogates; etc. Open domain, however domain knowledge proves important (e.g., scientific domain) Document Understanding Conferences (since 2000) contributed with much development in the field From 2008 the Text Analysis Conference Integration of technologies for background gathering (Gaizauskas&al07) Background gathering: the task of collecting information from the news wire and other archives to contextualise and support a breaking news story Backgrounder components • • Collaboration with Press Association • • similar events in the past; role players’ profiles; factual information on the event 11 year archive with more than 8 million stories Information access system comprising: Information Retrieval, Text Summarization, Question Answering, “Similar Event Search” Background Examples Breaking News “Powerful earthquake shook Turkey today” Past Similar Events “Last year an earthquake measuring 6.3-magnitude hit southern Turkey killing 144 people.” Extremes “Europe's biggest quake hit Lisbon, Portugal, on November 1, 1755, when 60,000 people died as the city was devastated and giant waves 10 metres high swept through the harbour and on to the shore.” Definitions “Quakes occur when the Earth's crust fractures, a process that can be caused by volcanic activity, landslides or subterranean collapse. The resulting plates grind together causing the tremors. Text Analysis Resources General Architecture for Text Engineering SUPPLE Parser (http://nlp.shef.ac.uk/research/supple) syntactic parsing and creation of logical forms Summarization Toolkit (http://gate.ac.uk) Tokenisation, Sentence Identification, POS tagging, NE recognition, etc. (http://www.dcs.shef.ac.uk/~saggion) Single and multi document summarization Lucene (http://lucene.apache.org) Text indexing and retrieval Summarization System Scores sentences based on numeric features in both single and multi-document cases position of sentence, similarity to headline, similarity to cluster centroid, etc. values are combined to obtain the sentence score Sentences represented as terms of vectors used to compute similarities (e.g. cosine metric) Redundancy detection for MDS is based on ngram overlap measure Press Association profiles are automatically identified Other profiles created using QA/summarization techniques Question Answering Passage (i.e., paragraph) retrieval using question Question and passage analysis using SUPPLE and discourse interpretation (i.e., coreference) expected answer type identified (EAT) Answer candidates in passages scored using sentence score (overlap with question) similarity of candidate answer to EAT (using WordNet) count relations between candidate and question entities merge scores across passages Semantic Representations Leading paragraphs are parsed using SUPPLE and semantic representations created “The head of Australia's biggest bank resigned today” head(e2), name(e4,'Australia'), country(e4), of(e3,e4), bank(e3), adj(e3,biggest), of(e2,e3), resign(e1), lsubj(e1,e2) Database records are created and used to support similar event search We can search for “resignation” events “resign” -> leave job = quit, renounce, leave office “head” -> person in charge = chief Finding Stories auto summaries profiles metadata stories Getting Answers answers context Getting Similar Events “jet dropped bomb in Iraq” jets drop bombs bombs dropped Extracting information for business intelligence applications MUSING Project – 6FP European Commission (ICT) integration of natural language processing and ontologies for business intelligence applications extraction of company information extraction of country/region information identification of opinions in text for company reputation Ontology-based IE in MUSING DATA SOURCE PROVIDER DOMAIN EXPERT DOCUMENT ONTOLOGY CURATOR USER MUSING ONTOLOGY DOCUMENT COLLECTOR USER INPUT DOCUMENT MUSING DATA REPOSITORY DOMAIN EXPERT MUSING APPLICATION ONTOLOGY-BASED DOCUMENT ANNOTATION ANNOTATED DOCUMENT ECONOMIC INDICATORS REGION SELECTION MODEL COMPANY INFORMATION ONTOLOGY POPULATION ENTERPRISE INTELLIGENCE REPORT KNOWLEDGE BASE INSTANCES & RELATIONS REGION RANK Data Sources in MUSING Data sources include balance sheets, company profiles, press data, web data, etc. (some private data) News papers from Italian financial news provider Companies’ web pages (main, “about us”, “contact us”, etc.) Wikipedia, CIA Fact Book, etc. Ontology is manually developed through interaction with domain experts and ontology curators It extends the PROTON ontology and covers the financial, international, and IT operational risk domain Company Information in MUSING Extracting Company Information Extracting information about a company requires for example identify the Company Name; Company Address; Parent Organization; Shareholders; etc. These associated pieces of information should be asserted as properties values of the company instance Statements for populating the ontology need to be created ( “Alcoa Inc” hasAlias “Alcoa”; “Alcoa Inc” hasWebPage “http://www.alcoa.com”, etc.) General Architecture for Text Engineering – GATE (Cunningham&al’02) Framework for development and deployment of natural language processing applications A graphical user interface allows users (computational linguists) access, composition and visualisation of different components and experimentation A Java library (gate.jar) for programmers to implement and pack applications Component Model Language Resources (LR) data Processing Resources (PR) algorithms Visualisation Resources (VR) graphical user interfaces (GUI) Components are extendable and user-customisable for example adaptation of an information extraction application to a new domain to a new language where the change involves adaptation of a module for word recognition and sentence recognition Documents in GATE A document is created from a file located somewhere in your disk or in a remote place or from a string A GATE document contains the “text” of your file and sets of annotations When the document is created and if a format analyser for your type is available “parsing” (format) will be applied and annotations will be created xml, sgml, html, etc. Documents also store features, useful for representing metadata about the document some features are created by GATE Documents in GATE Annotations have types (e.g. Token) belong to particular annotation sets start and end offsets – where in the document features and values which are used to store orthographic, grammatical, semantic information, etc. Documents can be grouped in a Corpus Corpus is other language resource in GATE which implements a set of documents Documents in GATE names in text semantics information Annotation Guidelines People need clear definition of what to annotate in the documents, with examples Typically written as a guidelines document Piloted first with few annotators, improved, then “real” annotation starts, when all annotators are trained Annotation tools require the definition of a formal DTD (e.g. XML schema) What annotation types are allowed What are their attributes/features and their values Optional vs obligatory; default values Annotation Schemas <?xml version="1.0"?> <schema xmlns="http://www.w3.org/2000/10/XMLSchema"> <!-- XSchema definition for email--> <element name="Email" /> </schema> Annotation Schemas <?xml version="1.0"?> <schema xmlns="http://www.w3.org/2000/10/XMLSchema"> <!-- XSchema definition for token--> <element name="Address"> <complexType> <attribute name="kind" use="optional"> <simpleType> <restriction base="string"> <enumeration value="email"/> <enumeration value="url"/> <enumeration value="phone"/> <enumeration value="ip"/> <enumeration value="street"/> <enumeration value="postcode"/> <enumeration value="country"/> <enumeration value="complete"/> </restriction> … Manual Annotation in GATE GUI Annotation in GATE GUI The following tasks can be carried out manually in the GATE GUI: Adding annotation sets Adding annotations Resizing them (changing boundaries) Deleting Changing highlighting colour Setting features and their values Preserving and exporting results Annotations can be stored as stand-off markup or in-line annotations The default method is standoff markup, where the annotations are stored separately from the text, so that the original text is not modified A corpus can also be saved as a regular or searchable (indexed) datastore Corpora and System Development “Gold standard” data created by manual annotation Corpora are divided typically into a training, sometimes testing, and unseen evaluation portion Rules and/or ML algorithms developed on the training part Tuned on the testing portion in order to optimise Rule priorities, rules effectiveness, etc. Parameters of the learning algorithm and the features used Evaluation set – the best system configuration is run on this data and the system performance is obtained No further tuning once evaluation set is used! Applications in GATE Applications are created by sequencing processing resources Applications can be run over a Corpus of documents – corpus pipeline so each component is applied to each document in the corpus in sequence Applications may not have a corpus as input, but different parameters – pipeline Name Entity Recognition Text Processing Tools Document Structure Analysis different document parsers take care of the structure of your document (xml, html, etc.) Tokenisation Sentence Identification Parts of speech tagging Morphological analysis All these language resources have as runtime parameter a GATE document, and they will produce annotations over it Most resources have initialisation parameters Creole a Collection of REusable Objects for Language Engineering ~ Language Resources + Processing Resources creole.xml provides details about available components, the java class that implements the resource, and jar file where it is found Example of resource <RESOURCE> </RESOURCE> <NAME>ANNIE English Tokeniser</NAME> <CLASS>gate.creole.tokeniser.DefaultTokeniser</CLASS> <COMMENT> A customisable English tokeniser (http://gate.ac.uk/sale/tao/#sec:en-tokeniser). </COMMENT> <PARAMETER NAME="document" COMMENT="The document to be tokenised" RUNTIME="true"> gate.Document </PARAMETER> <PARAMETER NAME="annotationSetName" RUNTIME="true" COMMENT="The annotation set to be used for the generated annotations" OPTIONAL="true"> java.lang.String </PARAMETER> <PARAMETER NAME="tokeniserRulesURL" DEFAULT="resources/tokeniser/DefaultTokeniser.rules" COMMENT="The URL for the rules file" SUFFIXES="rules"> java.net.URL </PARAMETER> <PARAMETER NAME="transducerGrammarURL" DEFAULT="resources/tokeniser/postprocess.jape" COMMENT="The URL for the postprocessing transducer" SUFFIXES="jape"> java.net.URL </PARAMETER> <PARAMETER NAME="encoding" COMMENT="The encoding used for reading the definitions" DEFAULT="UTF-8"> java.lang.String </PARAMETER> <ICON>tokeniser</ICON> Tokenisation Identify different words in text: numbers, symbols, words, etc. not only sequences between spaces or separators (in English) 2,000 is not “2” “,” “000” I’ve is “I” and “’ve” $20 is “$” and “20” Tokenisation in GATE Rule-based LHS > RHS LHS is a regular expression over character classes RHS specifies annotation to be created and features to be asserted for the created annotation "DECIMAL_DIGIT_NUMBER"+ >Token;kind=number; (SPACE_SEPARATOR) >SpaceToken;kind=space; (CONTROL) >SpaceToken;kind=control; "UPPERCASE_LETTER" (LOWERCASE_LETTER (LOWERCASE_LETTER|DASH_PUNCTUATION|FORMAT)*) * > Token;orth=upperInitial;kind=word; Tokeniser produces a Token type of annotation Tokenisation in GATE Features produced by the tokeniser string: the actual string of the token orth: orthographic information length: the length of the string kind: the type of token (word, symbol, punctuation, number) SpaceToken is another annotation produced features: kind (control or space), length, string Sentence Splitter end of sentence not end of sentence Decide where a sentence ends A rule based mechanism “The court ruled that Dr. Smith was innocent.” uses a list of known abbreviations to help identify end of sentence uses (e.g. Dr) a period is a sentence break if it is preceded by a non-abbreviation and followed by an uppercase common word The Splitter in GATE produces a Sentence type of annotation Parts of Speech Tagging Associate a part of speech tag to each word (Hepple’00) tags from the Penn Treebank including punctuation Based on Brill’s tagger but a different learning approach used for rule acquisition – no need for re-annotating the corpus at each iteration during learning Two steps during tagging initial guess based on lexicon (contain most likely tag) correction based on a list of rules (contextual) POS tags used CC - coordinating conjunction: ”and”, ”but”, ”nor”, ”or”, ”yet”, plus, minus, less, times (multiplication), over (division). Also ”for” (because) and ”so” (i.e., ”so that”). CD - cardinal number DT - determiner: Articles including ”a”, ”an”, ”every”, ”no”, ”the”, ”another”, ”any”, ”some”, ”those”. EX - existential there: Unstressed ”there” that triggers inversion of the inflected verb and the logical subject; ”There was a party in progress”. FW - foreign word IN - preposition or subordinating conjunction JJ - adjective: Hyphenated compounds that are used as modifiers; happy-go-lucky. JJR - adjective - comparative: Adjectives with the comparative ending ”-er” and a comparative meaning. Sometimes ”more” and ”less”. JJS - adjective - superlative: Adjectives with the superlative ending ”-est” (and ”worst”). Sometimes ”most”and ”least”. JJSS - -unknown-, but probably a variant of JJS -LRB- - -unknownLS - list item marker: Numbers and letters used as identifiers of items in a list. MD - modal: All verbs that don’t take an ”-s” ending in the third person singular present: ”can”, ”could”, ”dare”, ”may”, ”might”, ”must”, ”ought”, ”shall”, ”should”, ”will”, ”would”. NN - noun - singular or mass NNP - proper noun - singular: All words in names usually are capitalized but titles might not be. NNPS - proper noun - plural: All words in names usually are capitalized but titles might not be. NNS - noun - plural PDT - predeterminer: Determinerlike elements preceding an article or possessive pronoun; ”all/PDT his marbles”, ”quite/PDT a mess”. POS - possesive ending: Nouns ending in ”’s” or ”’”. PP - personal pronoun PRPR$ - unknown-, but probably possessive pronoun PRP - unknown-, but probably possessive pronoun PRP$ - unknown, but probably possessive pronoun,such as ”my”, ”your”, ”his”, ”his”, ”its”, ”one’s”, ”our”, and ”their”. RB - adverb: most words ending in ”-ly”. Also ”quite”, ”too”, ”very”, ”enough”, ”indeed”, ”not”, ”-n’t”, and ”never”. RBR - adverb - comparative: adverbs ending with ”-er” with a comparative meaning. RBS - adverb - superlative RP - particle: Mostly monosyllabic words that also double as directional adverbs. STAART - start state marker (used internally) SYM - symbol: technical symbols or expressions that aren’t English words. TO - literal to UH - interjection: Such as ”my”, ”oh”, ”please”, ”uh”, ”well”, ”yes”. VBD - verb - past tense: includes conditional form of the verb ”to be”; ”If I were/VBD rich...”. VBG - verb - gerund or present participle VBN - verb - past participle VBP - verb - non-3rd person singular present VB - verb - base form: subsumes imperatives, infinitives and subjunctives. VBZ - verb - 3rd person singular present WDT - wh-determiner WP$ - possesive wh-pronoun: includes ”whose” WP - wh-pronoun: includes ”what”, ”who”, and ”whom”. WRB - wh-adverb: includes ”how”, ”where”, ”why”. Includes ”when” when used in a temporal sense. Parts of Speech Tagging Two resources lexicon collected from corpus with <word, list of valid tags> some heuristics for unknown words rules for correcting tagging mistakes NN VBG PREVWD before rules instantiate patterns such as: employs VBZ empty JJ VB VBP Change tag A to tag B if Condition The GATE tagger produces a feature category for each token in the document, the value of the feature is the name of the POS tag Morphological Analysis in GATE For each noun and verb in the document identifies lemma and affix which are stored in the Token annotation (‘root’, ‘affix’) A set of rules for regular cases is used A set of irregular cases which explicitly indicate how to decompose the word is also used Stemming in GATE Removing prefixes and suffixed of a word produces a feature ‘stem’ in the Token annotation John -> stem=john tells -> stem=tell considered -> stem=consid (root=consider) leaving -> stem=leav (root=leave) had -> stem=had (root=have) Available for English and other languages (e.g. Spanish) Named Entity Recognition It is the cornerstone of many NLP applications – in particular of IE Identification of named entities in text Classification of the found strings in categories or types General types are Person Names, Organizations, Locations Others are Dates, Numbers, e-mails, Addresses, etc. Domains may have specific NEs: film names, drug names, programming languages, names of proteins, etc. NER problems There are problems even with well known categories “Ambrose Chapel” it’s not a name it is a place!!!! Ambiguity is one problem “Paris” can be a city or a person Paris (for Paris Hilton, the Person); Paris Hilton hotel (the place) London can be a place or an organization (the government) Approaches to NER Knowledge-based approach Word level information is useful in recognising entities: capitalization, type of word (number, symbol) Specialized lexicons (Gazetteer lists) usually created by hand; although methods exist to compile them from corpora List of known continents, countries, cities, person first names On-line resources are available to pull out that information Approaches to NER In GATE Gazetteers lists entries may contain some useful semantic information for example one may associate some features and values to entry names features can be used in grammars or can be used to enrich system output gazetteer lists are organized in index files Approaches to NER Knowledge-based approach Rules are used to combine different evidences a known first name followed by a sequence of words with upper initial may indicate a person name a upper initial word followed by a company designator (e.g., Co., Ltd.) may indicate a company name A cascade approach is generally used where some basic names are first identified and are latter combined into more complex names Gazetteers in GATE Lists store keywords (one keyword per line) list of male names (person_male.lst) Aaron Abraham …. Set of lists compiled and a finite state machine is created which operates on the strings The machine produces annotations of type Lookup when the keyword is found in text 60k entries in 80 types: organization; artifact; location; amount_unit; manufacturer; 60 Gazetteer in GATE Sets of lists are organized in a main lists file Each list specifies attributes MajorType and MinorType and Language, having major and minor types gives some flexibility to grammar rules government.lst:organization:government department.lst:organization:department person_male.lst:person_first:male person_female.lst:person_first:female Attributes are used to help identification of more complex entities (for example discriminatining when possible between a male or female name) List entries may be entities or parts of entities, or they may contain contextual information (e.g. job titles often indicate people) 61 Named Entity Grammar in GATE Implemented in the JAPE language (part of GATE) Regular expressions over annotations Provide access and manipulation of annotations produced by other modules Rules are stored in grammar files Grammar files are compiled into Finite State Machines A main grammar files specifies how different grammars should be executed (phases) constitute a cascade of FSTs over annotations 62 NER in GATE Rules are hand-coded, so some linguistic expertise is needed here uses annotations from tokeniser, POS tagger, and gazetteer modules use of contextual information rule priority based on pattern length, rule status and rule ordering Common entities: persons, locations, organisations, dates, addresses. 63 JAPE Language A JAPE grammar rule consists of a left hand side (LHS) and a right hand side (RHS) LHS= what to match (the pattern) RHS = how to annotate the found sequence A JAPE grammar is a sequence of grammar rules Grammars are compiled into finite state machines Rules have priority (number) There is a way to control how to match options parameter in the grammar files 64 LHS of JAPE rules The LHS of the rule contains patterns to be matched, in the form of annotations (and optionally their attributes). Annotation types to be recognized must be declared at the beginning of the phase Annotations may be combined using traditional operators [ | * + ?] LHS of JAPE rules There is no negative operator More than one pattern can be matched in a single rule Left and right context (not to be annotated) can be matched LHS has labels to be referred to in RHS Examples of LHS patterns //identify a token with upper initial ({Token.orth == upperInitial}):upper //recognise a sequence of one upper initial word followed by a location designator (e.g. Ennerdale Lake) ({Token.orth == upperInitial} {Lookup.majorType == loc_designator}) :location //same but with upper initial or all capitals (({Token.orth == upperInitial}|{Token.orth == allCaps}) {Lookup.majorType == loc_designator}) :location Example of RHS ({Token.orth == upperInitial} {Lookup.majorType == lake_designator}) :location :location.Location = { type = “lake” } Indicates annotation type to be produced ‘Location’ and features and values for that annotation type Macros in JAPE grammars Macro: ONE_DIGIT ({Token.kind == number, Token.length == "1"}) Macro: TWO_DIGIT ({Token.kind == number, Token.length == "2"}) Macro: FOUR_DIGIT ({Token.kind == number, Token.length == "4"}) Macro: DAY_MONTH_NUM (ONE_DIGIT | TWO_DIGIT) In the LHS of the rule one can use the macro name: (DAY_MONTH_NUM):annotate -> :annotate.DAY = {} Example of RHS (context) Rule: Date ({Token.string == "Date"} {Token.string == ":"}):context (({Token.kind == "number", Token.length == "2"}) ({Token}) ({Token.kind == "number", Token.length == "2"}) ({Token}) ({Token.kind == "number", Token.length == "2"})):annotate --> :annotate.Date = { type = "dd/mm/yy format" } Further processing in RHS Java code can be included in the RHS of the rule It is a powerful mechanism which can help add semantic information to the annotations for example extracting information from the context Available Java objects bindings: labels used in LHS are available doc: the GATE document which is being process annotations: all GATE document annotations produced until that stage inputAS, outputAS: phase input and output annotations JAPE Application modes Matching control for rules Brill (fires all matches) First (shortest match fires) Once (Phase exits after first match) All (as for Brill, but matching continues from offset following the current one, not from the end of the last match) Appelt (priority ordering: longest match fires, then explicit rule priority, then first defined rule fires) 73 JAPE Application Modes A A • {A}+ A Appelt Once First Brill All 75 Using phases Grammars usually consist of several phases, run sequentially A definition phase (conventionally called main.jape) lists the phases to be used, in order Only the definition phase needs to be loaded Temporary annotations may be created in early phases and used as input for later phases Annotations from earlier phases may need to be combined or modified Coreference Resolution Name coreference matches similar names in text, e.g. “Dr. Jacob Smith” and “Smith” creates a ‘matches’ annotation which allows you to extract a chain of ‘equivalent’ names Pronominal coreference solves references to named entities of pronouns in English (tokens marked with POS category PRP or PRP$) Coreference Resolution Orthographic co-reference can improve NE results by assigning entity type to previously unclassified names, based on relations with classified NEs May not reclassify already classified entities Classification of unknown entities very useful for surnames which match a full name, or abbreviations, e.g. “Bonfield” will match “Sir Peter Bonfield” ANNIE System A Nearly New Information Extraction System recognizes named entities in text “packed” application combining/sequencing the following components: document reset, tokeniser, splitter, tagger, gazetteer lookup, NE grammars, name coreference can be used as starting point to develop a new named recogniser Some NE Annotated Corpora MUC-6 and MUC-7 corpora - English CONLL shared task corpora http://cnts.uia.ac.be/conll2003/ner/ - NEs in English and German http://cnts.uia.ac.be/conll2002/ner/ - NEs in Spanish and Dutch TIDES surprise language exercise (NEs in Cebuano and Hindi) ACE – English http://www.ldc.upenn.edu/Projects/ACE/ The MUC-7 corpus 100 documents in SGML News domain 1880 Organizations (46%) 1324 Locations (32%) 887 Persons (22%) Inter-annotator agreement very high (~97%) http://www.itl.nist.gov/iaui/894.02/related_proj ects/muc/proceedings/muc_7_proceedings/m arsh_slides.pdf The MUC-7 Corpus (2) <ENAMEX TYPE="LOCATION">CAPE CANAVERAL</ENAMEX>, <ENAMEX TYPE="LOCATION">Fla.</ENAMEX> &MD; Working in chilly temperatures <TIMEX TYPE="DATE">Wednesday</TIMEX> <TIMEX TYPE="TIME">night</TIMEX>, <ENAMEX TYPE="ORGANIZATION">NASA</ENAMEX> ground crews readied the space shuttle Endeavour for launch on a Japanese satellite retrieval mission. <p> Endeavour, with an international crew of six, was set to blast off from the <ENAMEX TYPE="ORGANIZATION|LOCATION">Kennedy Space Center</ENAMEX> on <TIMEX TYPE="DATE">Thursday</TIMEX> at <TIMEX TYPE="TIME">4:18 a.m. EST</TIMEX>, the start of a 49minute launching period. The <TIMEX TYPE="DATE">nine day</TIMEX> shuttle flight was to be the 12th launched in darkness. Performance Evaluation Evaluation metric – mathematically defines how to measure the system’s performance against a human-annotated, gold standard Scoring program – implements the metric and provides performance measures For each document and over the entire corpus For each type of NE The Evaluation Metric Precision = correct answers/answers produced Recall = correct answers/total possible correct answers Trade-off between precision and recall F-Measure = (β2 + 1)PR / β2R + P [van Rijsbergen 75] β reflects the weighting between precision and recall, typically β=1 The GATE Evaluation Tool Regression Testing Need to track system’s performance over time When a change is made to the system we want to know what implications are over the entire corpus Why: because an improvement in one case can lead to problems in others GATE offers automated tool to help with the NE development task over time Document Indexing Indexing with Lucene Populate a corpus Create a Data Store (java serialisation) Save corpus to DS the corpus become “indexable” by Lucene Index the corpus using document content To search use the Information Retrieval plug-in SearchPR processing resource (put it in a “pipeline”) Specify parameters of search (corpus, query, etc.) and run double clicking on SearchPR displays the results Annotations in Context (ANNIC) Create a linguistic/semantic index Create a Lucene Searchable DS Populate a corpus Apply ANNIE to the corpus Save corpus to DS Search in Context in the DS GUI Machine Learning Approach Given a corpus annotated with named entities we want to create a classifier which decides if a string of text is a NE or not …<person>Mr. John Smith</person>… …<date>16th May 2005</date> The problem of recognising NEs can be seen as a classification problem Machine Learning Approach Each named entity instance is transformed for the learning problem The problem is transformed in a binary classification problem …<person>Mr. John Smith</person>… Mr. is the beginning of the NE person Smith is the end of the NE person is token begin of NE person? is token end of NE person? Context is used as features for the classifier Parsing with SUPPLE (Gaizauskas&al’05) Sheffield University Prolog Parser for Language Engineering A bottom-up parser for English which produces syntactic and semantic sentence respresentations An attribute-value context-free grammar of English is used to derive syntactic representations (it includes a question grammar for QA applications) categories in the grammar have attributes and values which can be instantiated during parsing Parsing with SUPPLE The grammar covers the following constituents prepositional phrases; noun phrases; core verbs; verb phrases; relative clauses; sentences; questions The input to the parsing process is a ‘chart’ where both lexical items and multiword expressions (named entities) are allowed The output is the best possible ‘parse’ of the sentence, this can be partial Parsing with SUPPLE Semantics is constructed compositionally as the sentence is parsed nouns and verbs are represented as normalised unary predicates (cat, eat, etc.) Identifiers (ei) are used to refer to an entity or an event and are produced for each noun and verb cat(e1), eat(e2) binary predicates represent relations or attribute values of the entities or events; they are a fixed inventory used to represent grammatical and semantic relations lsubj(X,Y), lobj(X,Z), of(X,Y), name(X,Z),… Parsing with SUPPLE Example Tony Blair meets U.S. President Bush. identifies Tony Blair and Bush as Person type and U.S. is a Location type wraps those constituents so that SUPPLE does not have to analyse them rest of elements in the sentence are passed as words with POS, roots, number, gender, etc. Parsing with SUPPLE Syntactic Annotation (string) best_parse=( s ( np ( bnp ( bnp_core ( bnp_head ( ne_np ( sem_cat "Tony Blair" ) ) ) ) ) ) ( fvp ( vp ( vpcore ( fvpcore ( nonmodal_vpcore ( nonmodal_vpcore1 ( vpcore1 ( av ( v "meets" ) ) ) ) ) ) ) ( np ( bnp ( bnp_core ( premods ( premods ( premod ( ne_np ( sem_cat "U.S." ) ) ) ) ( premod ( ne_np ( names_np ( pn "President" ) ) ) ) ) ( bnp_head ( ne_np ( sem_cat "Bush" ))))))))) Parsing with SUPPLE Parsing with SUPPLE Semantic Annotation (array of strings) qlf=[name(e2,'Tony Blair'), person(e2), realisation(e2,offsets(0,10)), meet(e1), time(e1,present), aspect(e1,simple), voice(e1,active), lobj(e1,e3), name(e3,'Bush'), person(e3), name(e4,'U.S.'), location(e4), country(e4), realisation(e4,offsets(17,21)), qual(e3,e4), ne_tag(e5,offsets(22,31)), name(e5,'President'), realisation(e5,offsets(22,31)), qual(e3,e5), realisation(e3,offsets(17,36)), realisation(e1,offsets(11,36)), lsubj(e1,e2)] Parsing with SUPPLE A wrapper is provided in GATE given a text which has been POS-tagged and Morphologically analysed, maps the tokens in each sentence to the input expected by SUPPLE read the syntactic and semantic information from files and stores the information into the GATE documents as parse, semantics, syntax tree nodes Can be run with SICStus prolog, SWI prolog, and PrologCafe (Java implementation)